Skip to main content

AWS Trainium (First Generation)

Overview​

AWS Trainium is an AI training ASIC designed by Amazon Annapurna Labs, launched in 2021. The first generation is available through Trn1 instances with 16 Trainium1 chips per instance. Each chip contains 32 GB HBM and delivers 191 TFLOPS FP8 compute. The Trn1 32xlarge integrates 512 NeuronCore v2 cores.

Core Specifications​

ItemSpec
ArchitectureTrainium (first generation)
NeuronCore32 per chip (NeuronCore-v2)
Memory32 GB HBM
Memory Bandwidth0.8 TB/s
FP8 Compute191 TFLOPS
BF16/FP16191 TFLOPS
FP3248 TFLOPS
TDP~350 W
NeuronLinkNeuronLink-v2
Form FactorTrn1 instances (AWS)

Vendor Information​

ItemDetail
ManufacturerAmazon (Annapurna Labs)
Websitehttps://aws.amazon.com/machine-learning/trainium/
Product Pagehttps://aws.amazon.com/ec2/instance-types/trn1/
SDKhttps://awsdocs-neuron.readthedocs-hosted.com/

Key Features​

  • AWS Neuron SDK: PyTorch, JAX integration
  • EFA interconnect: Low-latency cluster networking
  • AWS EC2 only

Use Cases​

  • LLM training (Trn1 32xlarge)
  • Recommendation systems
  • Large-scale distributed training